I wanted to respond to confirm that not only does peak-detect mode affect how the digital channels are acquired, it affects how the current contents of the digital channels in acquisition memory are displayed.
The pulses I'm looking at are roughly 100ns in duration, so while short in comparison to my buffer length (20MS) they are still long compared to my acquisition rate (62.5Msps, 16ns sample period). As a result the pulses are always detected and present in acquisition memory regardless of the selected acquisition mode. They're just not always visible in the traces when zoomed out.
However, if for example I acquire some data in single-sample mode, stop the acquisition and then select peak-detect mode, the missing pulses appear in all my trace data. Additionally, if I drop the acquisition sample rate down to something like 1MHz I can still detect every 100ns pulse using peak-detect mode.
While IMO the digital traces should follow the convention of most other logic analyzers I've encountered and enforce a minimum width for state transitions, switching to peak-detect mode works, even on previously acquired data.